@belkadan Thoughts?

Okay, let's see. I ignored this when Michael first filed it because I figured it was intractable or at least impractical, but Clang being able to do this even in the presence of modules means that a bunch of the worst problems have been solved. Let's take just the simplest approach: a crash during type-checking or SIL optimization, where re-running the same compiler will very likely end up with the same crash. What information do we need to collect/modify?

• The command-line invocation of the frontend, obviously. Modifications: search paths? including those passed to Clang through -Xcc?

• Any filelists / output file maps used in that invocation. Modifications: these are usually absolute paths, so we need to change those

• All files in the current target.

• All imported Clang headers and module maps, obviously.

• All imported Swift modules. If something was imported using its module interface, we can grab that; if it was imported using a compiled module, we can grab that. (Ideally we'd get both; even when importing something using its interface, it'll have a compiled module form, either prebuilt next to the compiler or cached near the Clang PCMs in the module cache directory. These record dependency paths like PCMs do, though.)

I think that's all you'd need to handle regular frontend crashes. For crashes in the merge-modules phase, we'd go a bit further and add the partial modules into the mix, which shouldn't be hard because they're inputs of the merge-modules frontend invocation.

@belkadan Thanks for looking at this. Being able to do this automagically would make it significantly easier for us to improve the quality of the compiler and make our users happy that they do not need to "construct" a test case. That being said, I think we would need to have some sort of modal dialog for automatic sending (or maybe something that lets the user attach it to a radar). Perhaps we can obfuscate/minify (not sure).

I'm very concerned about obfuscation because (sadly) some parts of the compiler use user-written names for ordering, and therefore we might lose the bug. (Thinking of protocol ordering in particular, which is used for generic signature canonicalization.)

Ok. Even without the obfuscation I think it is still worth it.